Give ChatGPT a sample of your own writing and tell it to match that voice, then name the specific habits you want removed — that combination works far better than simply typing "make it sound more human."
A vague instruction leaves the model guessing, so it falls back on the same polished, slightly stiff register it uses by default. When you show it a paragraph you actually wrote and list the phrases you can't stand, you're replacing guesswork with a target it can hit.
The mechanism here is worth understanding, because it explains why most tone requests fail. ChatGPT doesn't have a single "human" setting buried in a menu. It predicts the most likely next words given everything in the conversation, and by default the most likely words for a professional-sounding request are the ones that appear most often in polished business writing — phrases like "leverage," "seamless," "in today's fast-paced world," and "it's important to note that."
Those aren't random. They're the statistical center of gravity for formal text. To shift the output, you have to shift the center of gravity by feeding in different examples and explicitly banning the patterns you don't want.
A style sample does more work than any adjective, because the model can copy rhythm, sentence length, and word choice directly rather than interpreting what "casual" means to you. According to our AI tool database, ChatGPT is OpenAI's flagship assistant, currently listed as GPT-5.5 with a 1M-token context window, which matters here: a large context window means you can paste a full sample of your writing — several paragraphs, even a whole previous article — without it getting truncated. That's a practical advantage for tone matching that smaller context models can't offer.
Here's a concrete worked example. Suppose you want a short update for your team and ChatGPT keeps producing this: "I am writing to inform you that the project timeline has been adjusted to accommodate recent developments." Paste two or three sentences of how you actually talk — something like "Quick heads up — we're pushing the launch back a week.
Nothing's broken, we just need more testing time" — and add an instruction: "Match this voice. Short sentences. No 'I am writing to inform you.'
No 'leverage.' Write like you'd explain it to a colleague over coffee." The output typically shifts to something like: "Quick update: launch is moving back a week.
Nothing's wrong — we just want more testing time before it goes out." Same information, different register. The tip that goes beyond the obvious: ask it to list the formal phrases it removed.
That gives you a reusable ban list you can paste into future prompts, so you're not re-teaching the same lesson every time. You can also ask it to rewrite your original sample in its default style first, so you can see exactly what you're correcting against — the contrast makes the instruction sharper.
Where this advice does not apply: regulated, legal, medical, or compliance copy. If you're writing terms of service, a safety notice, or anything a lawyer will read, informality isn't a style choice — it's a liability. Plain, slightly stiff language exists in those documents for a reason, and asking ChatGPT to "sound like a colleague" there is the wrong move.
The same goes for certain corporate communications where a neutral register signals respect or distance on purpose. Tone matching is a tool for blogs, emails, social posts, and internal updates — not a universal upgrade. It also costs you something: every style sample you paste eats context, and if you paste too many examples the model starts averaging them into a mush that sounds like nobody.
Two or three tight samples beat ten loose ones. For a related problem — catching output that still reads like a machine wrote it — see How do I stop AI from writing content that sounds like a robot wrote it?.